Inference float16 support

[kexinzhao]

Below are the list of to-do items for supporting float16 inference.

• make float16 a POD type (Make float16 a C++11 POD class #8456)
• data_type_transform for fp16 (Integrate float16 into data_type_transform #8619 Add context wait in data_type_transform #8850)
• add fp16 gemm function into math_function (Add float16 GEMM math function on GPU #8695)
• add fp16 support for mul_op (Add fp16 support for Mul Op #8817)
• bind float16 to numpy.float16
• add fp16 for cudnn conv2d_op
• add fp16 for cudnn pool2d_op
• cast op
• batch norm op
• dropout
• use tensor core for cublas
• use tensor core for cudnn
• have a vgg16 cifar10 inference python example that runs on float32 mode
• have a vgg16 cifar10 inference python example that runs on float16 mode
• get some initial result (verify tensor cores works, verify fp16 is faster, verify accuracy is comparable
• relu
• element-add op
• softmax
• tanh and other activations
• float16 example python
• float16 example c++ inference
• float16 example correctness
• float16 transpiler and example
• update float16 design doc
• save
• load
• float16 correctness code
• float16 benchmark code
• float16 report in English
• conv_3d
• pool_3d
• sum op
• other element-wise operator (sub divide mul min max pow)
• cudnn relu